Colorectal cancer is a major global health concern, where early detection of polyps through colonoscopy significantly reduces mortality. While recent deep learning-based segmentation methods achieve high accuracy, many rely on computationally expensive architectures that limit real-time deployment. This paper proposes EBA-Net, an Edge-Boundary Aware attention network designed to balance segmentation performance with computational efficiency. The model integrates a lightweight ResNet-18 encoder with multi-scale feature aggregation, an auxiliary edge detection branch, and a boundary-aware attention module that enhances feature representation in ambiguous regions. On the Kvasir-SEG benchmark, EBA-Net reaches an IoU of 0.881 and Dice score of 0.936 while using only 12.11 million parameters. This represents a 78% reduction in model size compared to recent state-of-the-art methods. These results suggest that careful attention to boundary features can deliver competitive accuracy without the computational overhead of larger models.

Polyp segmentation, deep learning, medical image analysis, attention mechanism, boundary detection, efficient neural networks